At CEA-List, I built ABACUS, a judge-free benchmark for numerical image editing that checks edits against panoptic ground truth. Its evaluation recovered the human count on 0.82 of edits, compared with 0.55 for a VLM judge, and the work was accepted at WACV 2027.
I also designed TALLY, a count-control head for frozen diffusion and flow editors. It ranked second of 12 on I2EBench and first on Counting.
At the Computer Vision Center (CVC) × CRNS, I developed collaborative, multi-turn and debate-based frameworks for agentic vision-language models in document understanding. This work resulted in a first-author paper accepted at CVPR 2026.
My research has also included a compact mixture-of-experts transformer for C code vulnerability detection at MBZUAI and causal structured graphs from interviews at Hamad Bin Khalifa University (QCRI). I’m a Kaggle Competitions Master, and I ranked 17th out of 85,000 participants on Zindi.
